322 lines · c
1// RUN: %clang_cc1 -no-enable-noundef-analysis -fenable-matrix -triple x86_64-apple-darwin %s -emit-llvm -disable-llvm-passes -o - | FileCheck --check-prefixes=COMMON,CHECK64 %s2// RUN: %clang_cc1 -no-enable-noundef-analysis -fenable-matrix -triple i386-apple-darwin %s -emit-llvm -disable-llvm-passes -o - | FileCheck --check-prefixes=COMMON,CHECK32 %s3 4// Also check we do not crash when running some middle-end passes. Most5// importantly this includes the IR verifier, to ensure we emit valid IR.6// RUN: %clang_cc1 -fenable-matrix -emit-llvm -triple x86_64-apple-darwin %s -o %t7 8// Tests for the matrix type builtins.9 10typedef double dx5x5_t __attribute__((matrix_type(5, 5)));11typedef float fx2x3_t __attribute__((matrix_type(2, 3)));12typedef float fx3x2_t __attribute__((matrix_type(3, 2)));13typedef int ix20x4_t __attribute__((matrix_type(20, 4)));14typedef int ix4x20_t __attribute__((matrix_type(4, 20)));15typedef unsigned ux1x6_t __attribute__((matrix_type(1, 6)));16typedef unsigned ux6x1_t __attribute__((matrix_type(6, 1)));17 18void transpose_double_5x5(dx5x5_t *a) {19 // COMMON-LABEL: define{{.*}} void @transpose_double_5x5(20 // CHECK32: [[A:%.*]] = load <25 x double>, ptr {{.*}}, align 421 // CHECK64: [[A:%.*]] = load <25 x double>, ptr {{.*}}, align 822 // COMMON-NEXT: [[TRANS:%.*]] = call <25 x double> @llvm.matrix.transpose.v25f64(<25 x double> [[A]], i32 5, i32 5)23 // CHECK32-NEXT: store <25 x double> [[TRANS]], ptr %a_t, align 424 // CHECK64-NEXT: store <25 x double> [[TRANS]], ptr %a_t, align 825 26 dx5x5_t a_t = __builtin_matrix_transpose(*a);27}28 29void transpose_float_3x2(fx3x2_t *a) {30 // COMMON-LABEL: define{{.*}} void @transpose_float_3x2(31 // COMMON: [[A:%.*]] = load <6 x float>, ptr {{.*}}, align 432 // COMMON-NEXT: [[TRANS:%.*]] = call <6 x float> @llvm.matrix.transpose.v6f32(<6 x float> [[A]], i32 3, i32 2)33 // COMMON-NEXT: store <6 x float> [[TRANS]], ptr %a_t, align 434 35 fx2x3_t a_t = __builtin_matrix_transpose(*a);36}37 38void transpose_int_20x4(ix20x4_t *a) {39 // COMMON-LABEL: define{{.*}} void @transpose_int_20x4(40 // COMMON: [[A:%.*]] = load <80 x i32>, ptr {{.*}}, align 441 // COMMON-NEXT: [[TRANS:%.*]] = call <80 x i32> @llvm.matrix.transpose.v80i32(<80 x i32> [[A]], i32 20, i32 4)42 // COMMON-NEXT: store <80 x i32> [[TRANS]], ptr %a_t, align 443 44 ix4x20_t a_t = __builtin_matrix_transpose(*a);45}46 47struct Foo {48 ux1x6_t in;49 ux6x1_t out;50};51 52void transpose_struct_member(struct Foo *F) {53 // COMMON-LABEL: define{{.*}} void @transpose_struct_member(54 // COMMON: [[M:%.*]] = load <6 x i32>, ptr {{.*}}, align 455 // COMMON-NEXT: [[M_T:%.*]] = call <6 x i32> @llvm.matrix.transpose.v6i32(<6 x i32> [[M]], i32 1, i32 6)56 // CHECK32-NEXT: [[F_ADDR:%.*]] = load ptr, ptr %F.addr, align 457 // CHECK64-NEXT: [[F_ADDR:%.*]] = load ptr, ptr %F.addr, align 858 // COMMON-NEXT: [[OUT_PTR:%.*]] = getelementptr inbounds nuw %struct.Foo, ptr [[F_ADDR]], i32 0, i32 159 // COMMON-NEXT: store <6 x i32> [[M_T]], ptr [[OUT_PTR]], align 460 61 F->out = __builtin_matrix_transpose(F->in);62}63 64void transpose_transpose_struct_member(struct Foo *F) {65 // COMMON-LABEL: define{{.*}} void @transpose_transpose_struct_member(66 // COMMON: [[M:%.*]] = load <6 x i32>, ptr {{.*}}, align 467 // COMMON-NEXT: [[M_T:%.*]] = call <6 x i32> @llvm.matrix.transpose.v6i32(<6 x i32> [[M]], i32 1, i32 6)68 // COMMON-NEXT: [[M_T2:%.*]] = call <6 x i32> @llvm.matrix.transpose.v6i32(<6 x i32> [[M_T]], i32 6, i32 1)69 // CHECK32-NEXT: [[F_ADDR:%.*]] = load ptr, ptr %F.addr, align 470 // CHECK64-NEXT: [[F_ADDR:%.*]] = load ptr, ptr %F.addr, align 871 // COMMON-NEXT: [[IN_PTR:%.*]] = getelementptr inbounds nuw %struct.Foo, ptr [[F_ADDR]], i32 0, i32 072 // COMMON-NEXT: store <6 x i32> [[M_T2]], ptr [[IN_PTR]], align 473 74 F->in = __builtin_matrix_transpose(__builtin_matrix_transpose(F->in));75}76 77dx5x5_t get_matrix(void);78 79void transpose_rvalue(void) {80 // COMMON-LABEL: define{{.*}} void @transpose_rvalue()81 // COMMON-NEXT: entry:82 // CHECK32-NEXT: [[M_T_ADDR:%.*]] = alloca [25 x double], align 483 // CHECK64-NEXT: [[M_T_ADDR:%.*]] = alloca [25 x double], align 884 // COMMON-NEXT: [[CALL:%.*]] = call <25 x double> @get_matrix()85 // COMMON-NEXT: [[M_T:%.*]] = call <25 x double> @llvm.matrix.transpose.v25f64(<25 x double> [[CALL]], i32 5, i32 5)86 // CHECK32-NEXT: store <25 x double> [[M_T]], ptr [[M_T_ADDR]], align 487 // CHECK64-NEXT: store <25 x double> [[M_T]], ptr [[M_T_ADDR]], align 888 89 dx5x5_t m_t = __builtin_matrix_transpose(get_matrix());90}91 92const dx5x5_t global_matrix;93 94void transpose_global(void) {95 // COMMON-LABEL: define{{.*}} void @transpose_global()96 // COMMON-NEXT: entry:97 // CHECK32-NEXT: [[M_T_ADDR:%.*]] = alloca [25 x double], align 498 // CHECK32-NEXT: [[GLOBAL_MATRIX:%.*]] = load <25 x double>, ptr @global_matrix, align 499 // CHECK64-NEXT: [[M_T_ADDR:%.*]] = alloca [25 x double], align 8100 // CHECK64-NEXT: [[GLOBAL_MATRIX:%.*]] = load <25 x double>, ptr @global_matrix, align 8101 // COMMON-NEXT: [[M_T:%.*]] = call <25 x double> @llvm.matrix.transpose.v25f64(<25 x double> [[GLOBAL_MATRIX]], i32 5, i32 5)102 // CHECK32-NEXT: store <25 x double> [[M_T]], ptr [[M_T_ADDR]], align 4103 // CHECK64-NEXT: store <25 x double> [[M_T]], ptr [[M_T_ADDR]], align 8104 105 dx5x5_t m_t = __builtin_matrix_transpose(global_matrix);106}107 108void column_major_load_with_const_stride_double(double *Ptr) {109 // COMMON-LABEL: define{{.*}} void @column_major_load_with_const_stride_double(ptr %Ptr)110 // CHECK32: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 4111 // CHECK32-NEXT: call <25 x double> @llvm.matrix.column.major.load.v25f64.i32(ptr align 4 [[PTR]], i32 5, i1 false, i32 5, i32 5)112 // CHECK64: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 8113 // CHECK64-NEXT: call <25 x double> @llvm.matrix.column.major.load.v25f64.i64(ptr align 8 [[PTR]], i64 5, i1 false, i32 5, i32 5)114 115 dx5x5_t m_a1 = __builtin_matrix_column_major_load(Ptr, 5, 5, 5);116}117 118void column_major_load_with_const_stride2_double(double *Ptr) {119 // COMMON-LABEL: define{{.*}} void @column_major_load_with_const_stride2_double(ptr %Ptr)120 // CHECK32: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 4121 // CHECK32-NEXT: call <25 x double> @llvm.matrix.column.major.load.v25f64.i32(ptr align 4 [[PTR]], i32 15, i1 false, i32 5, i32 5)122 // CHECK64: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 8123 // CHECK64-NEXT: call <25 x double> @llvm.matrix.column.major.load.v25f64.i64(ptr align 8 [[PTR]], i64 15, i1 false, i32 5, i32 5)124 125 dx5x5_t m_a2 = __builtin_matrix_column_major_load(Ptr, 5, 5, 2 * 3 + 9);126}127 128void column_major_load_with_variable_stride_ull_float(float *Ptr, unsigned long long S) {129 // COMMON-LABEL: define{{.*}} void @column_major_load_with_variable_stride_ull_float(ptr %Ptr, i64 %S)130 // CHECK32: [[S:%.*]] = load i64, ptr %S.addr, align 8131 // CHECK32-NEXT: [[STRIDE_TRUNC:%.*]] = trunc i64 [[S]] to i32132 // CHECK32-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 4133 // CHECK32-NEXT: call <6 x float> @llvm.matrix.column.major.load.v6f32.i32(ptr align 4 [[PTR]], i32 [[STRIDE_TRUNC]], i1 false, i32 2, i32 3)134 135 // CHECK64: [[S:%.*]] = load i64, ptr %S.addr, align 8136 // CHECK64-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 8137 // CHECK64-NEXT: call <6 x float> @llvm.matrix.column.major.load.v6f32.i64(ptr align 4 [[PTR]], i64 [[S]], i1 false, i32 2, i32 3)138 139 fx2x3_t m_b = __builtin_matrix_column_major_load(Ptr, 2, 3, S);140}141 142void column_major_load_with_stride_math_int(int *Ptr, int S) {143 // COMMON-LABEL: define{{.*}} void @column_major_load_with_stride_math_int(ptr %Ptr, i32 %S)144 // COMMON: [[S:%.*]] = load i32, ptr %S.addr, align 4145 // COMMON-NEXT: [[STRIDE:%.*]] = add nsw i32 [[S]], 32146 // CHECK32-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 4147 // CHECK32-NEXT: call <80 x i32> @llvm.matrix.column.major.load.v80i32.i32(ptr align 4 [[PTR]], i32 [[STRIDE]], i1 false, i32 4, i32 20)148 //149 // CHECK64-NEXT: [[STRIDE_EXT:%.*]] = sext i32 [[STRIDE]] to i64150 // CHECK64-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 8151 // CHECK64-NEXT: call <80 x i32> @llvm.matrix.column.major.load.v80i32.i64(ptr align 4 [[PTR]], i64 [[STRIDE_EXT]], i1 false, i32 4, i32 20)152 153 ix4x20_t m_c = __builtin_matrix_column_major_load(Ptr, 4, 20, S + 32);154}155 156void column_major_load_with_stride_math_s_int(int *Ptr, short S) {157 // COMMON-LABEL: define{{.*}} void @column_major_load_with_stride_math_s_int(ptr %Ptr, i16 signext %S)158 // COMMON: [[S:%.*]] = load i16, ptr %S.addr, align 2159 // COMMON-NEXT: [[S_EXT:%.*]] = sext i16 [[S]] to i32160 // COMMON-NEXT: [[STRIDE:%.*]] = add nsw i32 [[S_EXT]], 32161 // CHECK32-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 4162 // CHECK32-NEXT: %matrix = call <80 x i32> @llvm.matrix.column.major.load.v80i32.i32(ptr align 4 [[PTR]], i32 [[STRIDE]], i1 false, i32 4, i32 20)163 //164 // CHECK64-NEXT: [[STRIDE_EXT:%.*]] = sext i32 [[STRIDE]] to i64165 // CHECK64-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 8166 // CHECK64-NEXT: %matrix = call <80 x i32> @llvm.matrix.column.major.load.v80i32.i64(ptr align 4 [[PTR]], i64 [[STRIDE_EXT]], i1 false, i32 4, i32 20)167 168 ix4x20_t m_c = __builtin_matrix_column_major_load(Ptr, 4, 20, S + 32);169}170 171void column_major_load_array1(double Ptr[25]) {172 // COMMON-LABEL: define{{.*}} void @column_major_load_array1(ptr %Ptr)173 // CHECK32: [[ADDR:%.*]] = load ptr, ptr %Ptr.addr, align 4174 // CHECK32-NEXT: call <25 x double> @llvm.matrix.column.major.load.v25f64.i32(ptr align 4 [[ADDR]], i32 5, i1 false, i32 5, i32 5)175 176 // CHECK64: [[ADDR:%.*]] = load ptr, ptr %Ptr.addr, align 8177 // CHECK64-NEXT: call <25 x double> @llvm.matrix.column.major.load.v25f64.i64(ptr align 8 [[ADDR]], i64 5, i1 false, i32 5, i32 5)178 179 dx5x5_t m = __builtin_matrix_column_major_load(Ptr, 5, 5, 5);180}181 182void column_major_load_array2(void) {183 // COMMON-LABEL: define{{.*}} void @column_major_load_array2() #0 {184 // COMMON-NEXT: entry:185 // CHECK32-NEXT: [[PTR:%.*]] = alloca [25 x double], align 8186 // CHECK32: [[ARRAY_DEC:%.*]] = getelementptr inbounds [25 x double], ptr [[PTR]], i32 0, i32 0187 // CHECK32-NEXT: call <25 x double> @llvm.matrix.column.major.load.v25f64.i32(ptr align 8 [[ARRAY_DEC]], i32 5, i1 false, i32 5, i32 5)188 189 // CHECK64-NEXT: [[PTR:%.*]] = alloca [25 x double], align 16190 // CHECK64: [[ARRAY_DEC:%.*]] = getelementptr inbounds [25 x double], ptr [[PTR]], i64 0, i64 0191 // CHECK64-NEXT: call <25 x double> @llvm.matrix.column.major.load.v25f64.i64(ptr align 16 [[ARRAY_DEC]], i64 5, i1 false, i32 5, i32 5)192 193 double Ptr[25];194 dx5x5_t m = __builtin_matrix_column_major_load(Ptr, 5, 5, 5);195}196 197void column_major_load_const(const double *Ptr) {198 // COMMON-LABEL: define{{.*}} void @column_major_load_const(ptr %Ptr)199 // CHECK32: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 4200 // CHECK32-NEXT: call <25 x double> @llvm.matrix.column.major.load.v25f64.i32(ptr align 4 [[PTR]], i32 5, i1 false, i32 5, i32 5)201 //202 // CHECK64: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 8203 // CHECK64-NEXT: call <25 x double> @llvm.matrix.column.major.load.v25f64.i64(ptr align 8 [[PTR]], i64 5, i1 false, i32 5, i32 5)204 205 dx5x5_t m_a1 = __builtin_matrix_column_major_load(Ptr, 5, 5, 5);206}207 208void column_major_load_volatile(volatile double *Ptr) {209 // COMMON-LABEL: define{{.*}} void @column_major_load_volatile(ptr %Ptr)210 // CHECK32: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 4211 // CHECK32-NEXT: call <25 x double> @llvm.matrix.column.major.load.v25f64.i32(ptr align 4 [[PTR]], i32 5, i1 true, i32 5, i32 5)212 //213 // CHECK64: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 8214 // CHECK64-NEXT: call <25 x double> @llvm.matrix.column.major.load.v25f64.i64(ptr align 8 [[PTR]], i64 5, i1 true, i32 5, i32 5)215 216 dx5x5_t m_a1 = __builtin_matrix_column_major_load(Ptr, 5, 5, 5);217}218 219void column_major_store_with_const_stride_double(double *Ptr) {220 // COMMON-LABEL: define{{.*}} void @column_major_store_with_const_stride_double(ptr %Ptr)221 // CHECK32: [[M:%.*]] = load <25 x double>, ptr {{.*}}, align 4222 // CHECK32-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 4223 // CHECK32-NEXT: call void @llvm.matrix.column.major.store.v25f64.i32(<25 x double> [[M]], ptr align 4 [[PTR]], i32 5, i1 false, i32 5, i32 5)224 //225 // CHECK64: [[M:%.*]] = load <25 x double>, ptr {{.*}}, align 8226 // CHECK64-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 8227 // CHECK64-NEXT: call void @llvm.matrix.column.major.store.v25f64.i64(<25 x double> [[M]], ptr align 8 [[PTR]], i64 5, i1 false, i32 5, i32 5)228 229 dx5x5_t m;230 __builtin_matrix_column_major_store(m, Ptr, 5);231}232 233void column_major_store_with_const_stride2_double(double *Ptr) {234 // COMMON-LABEL: define{{.*}} void @column_major_store_with_const_stride2_double(ptr %Ptr)235 // CHECK32: [[M:%.*]] = load <25 x double>, ptr {{.*}}, align 4236 // CHECK32-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 4237 // CHECK32-NEXT: call void @llvm.matrix.column.major.store.v25f64.i32(<25 x double> [[M]], ptr align 4 [[PTR]], i32 15, i1 false, i32 5, i32 5)238 //239 // CHECK64: [[M:%.*]] = load <25 x double>, ptr {{.*}}, align 8240 // CHECK64-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 8241 // CHECK64-NEXT: call void @llvm.matrix.column.major.store.v25f64.i64(<25 x double> [[M]], ptr align 8 [[PTR]], i64 15, i1 false, i32 5, i32 5)242 //243 dx5x5_t m;244 __builtin_matrix_column_major_store(m, Ptr, 2 * 3 + 9);245}246 247void column_major_store_with_stride_math_int(int *Ptr, int S) {248 // COMMON-LABEL: define{{.*}} void @column_major_store_with_stride_math_int(ptr %Ptr, i32 %S)249 // COMMON: [[M:%.*]] = load <80 x i32>, ptr {{.*}}, align 4250 // CHECK32-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 4251 // CHECK64-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 8252 // COMMON-NEXT: [[S:%.*]] = load i32, ptr %S.addr, align 4253 // COMMON-NEXT: [[ADD:%.*]] = add nsw i32 [[S]], 32254 // CHECK32-NEXT: call void @llvm.matrix.column.major.store.v80i32.i32(<80 x i32> [[M]], ptr align 4 [[PTR]], i32 [[ADD]], i1 false, i32 4, i32 20)255 //256 // CHECK64-NEXT: [[IDX:%.*]] = sext i32 [[ADD]] to i64257 // CHECK64-NEXT: call void @llvm.matrix.column.major.store.v80i32.i64(<80 x i32> [[M]], ptr align 4 [[PTR]], i64 [[IDX]], i1 false, i32 4, i32 20)258 259 ix4x20_t m;260 __builtin_matrix_column_major_store(m, Ptr, S + 32);261}262 263void column_major_store_with_stride_math_s_int(int *Ptr, short S) {264 // COMMON-LABEL: define{{.*}} void @column_major_store_with_stride_math_s_int(ptr %Ptr, i16 signext %S)265 // COMMON: [[M:%.*]] = load <80 x i32>, ptr {{.*}}, align 4266 // CHECK32-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 4267 // CHECK64-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 8268 // COMMON-NEXT: [[S:%.*]] = load i16, ptr %S.addr, align 2269 // COMMON-NEXT: [[EXT:%.*]] = sext i16 [[S]] to i32270 // COMMON-NEXT: [[ADD:%.*]] = add nsw i32 [[EXT]], 2271 // CHECK32-NEXT: call void @llvm.matrix.column.major.store.v80i32.i32(<80 x i32> [[M]], ptr align 4 [[PTR]], i32 [[ADD]], i1 false, i32 4, i32 20)272 //273 // CHECK64-NEXT: [[IDX:%.*]] = sext i32 [[ADD]] to i64274 // CHECK64-NEXT: call void @llvm.matrix.column.major.store.v80i32.i64(<80 x i32> [[M]], ptr align 4 [[PTR]], i64 [[IDX]], i1 false, i32 4, i32 20)275 276 ix4x20_t m;277 __builtin_matrix_column_major_store(m, Ptr, S + 2);278}279 280void column_major_store_array1(double Ptr[25]) {281 // COMMON-LABEL: define{{.*}} void @column_major_store_array1(ptr %Ptr)282 // CHECK32: [[M:%.*]] = load <25 x double>, ptr {{.*}}, align 4283 // CHECK32-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 4284 // CHECK32-NEXT: call void @llvm.matrix.column.major.store.v25f64.i32(<25 x double> [[M]], ptr align 4 [[PTR]], i32 5, i1 false, i32 5, i32 5)285 //286 // CHECK64: [[M:%.*]] = load <25 x double>, ptr {{.*}}, align 8287 // CHECK64-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 8288 // CHECK64-NEXT: call void @llvm.matrix.column.major.store.v25f64.i64(<25 x double> [[M]], ptr align 8 [[PTR]], i64 5, i1 false, i32 5, i32 5)289 290 dx5x5_t m;291 __builtin_matrix_column_major_store(m, Ptr, 5);292}293 294void column_major_store_array2(void) {295 // COMMON-LABEL: define{{.*}} void @column_major_store_array2()296 // CHECK32: [[M:%.*]] = load <25 x double>, ptr {{.*}}, align 4297 // CHECK32-NEXT: [[PTR:%.*]] = getelementptr inbounds [25 x double], ptr %Ptr, i32 0, i32 0298 // CHECK32-NEXT: call void @llvm.matrix.column.major.store.v25f64.i32(<25 x double> [[M]], ptr align 8 [[PTR]], i32 5, i1 false, i32 5, i32 5)299 //300 // CHECK64: [[M:%.*]] = load <25 x double>, ptr {{.*}}, align 8301 // CHECK64-NEXT: [[PTR:%.*]] = getelementptr inbounds [25 x double], ptr %Ptr, i64 0, i64 0302 // CHECK64-NEXT: call void @llvm.matrix.column.major.store.v25f64.i64(<25 x double> [[M]], ptr align 16 [[PTR]], i64 5, i1 false, i32 5, i32 5)303 304 double Ptr[25];305 dx5x5_t m;306 __builtin_matrix_column_major_store(m, Ptr, 5);307}308 309void column_major_store_volatile(volatile double *Ptr) {310 // COMMON-LABEL: define{{.*}} void @column_major_store_volatile(ptr %Ptr) #0 {311 // CHECK32: [[M:%.*]] = load <25 x double>, ptr {{.*}}, align 4312 // CHECK32-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 4313 // CHECK32-NEXT: call void @llvm.matrix.column.major.store.v25f64.i32(<25 x double> [[M]], ptr align 4 [[PTR]], i32 5, i1 true, i32 5, i32 5)314 //315 // CHECK64: [[M:%.*]] = load <25 x double>, ptr {{.*}}, align 8316 // CHECK64-NEXT: [[PTR:%.*]] = load ptr, ptr %Ptr.addr, align 8317 // CHECK64-NEXT: call void @llvm.matrix.column.major.store.v25f64.i64(<25 x double> [[M]], ptr align 8 [[PTR]], i64 5, i1 true, i32 5, i32 5)318 319 dx5x5_t m;320 __builtin_matrix_column_major_store(m, Ptr, 5);321}322